The Future of AI Startups Isn't Proprietary AI Models — It's Strategic Execution & Deep Domain Insights

In today's venture ecosystem, "proprietary" has become a buzzword almost detached from meaning.

Custom models, exclusive datasets, in-house architectures — founders race to showcase technical assets under the assumption that ownership automatically translates to value.

But when you peel back the layers, a different story often emerges: bloated teams, bloated R&D spend, and AI products that underperform cheaper, faster, open-source alternatives.

I am building an AI company and we use foundational models for some of our features. But do they devalue us? Absolutely not. We create with nuance, domain insight applied to a problem we deeply understand, aimed at a customer we deeply understand.

We think differently about proprietary models because we believe the future of successful AI startups won't be decided by who builds the largest model. It will be decided by who allocates capital intelligently, understands users deeply, and executes ruthlessly at the application layer.

Proprietary technology can be an advantage — when it’s strategic. But in most cases today, it’s a very expensive distraction.

The New Reality: Proprietary ≠ Progress

It’s easy to throw millions at building custom large language models. It’s harder to admit when those models actually underperform off-the-shelf foundational models like OpenAI's GPT-4, Anthropic's Claude, or Meta's Llama 3.

Open-source AI and API-accessible models have shifted the ground under our feet:

Massive economies of scale now favour foundational model providers. For most start-ups its insane to try and build something better or bigger. Open ecosystems are iterating faster than any single R&D team can manage.

Training your own model rarely delivers a step-change in capability — but almost always creates significant cost and maintenance burdens.

Yet fundraising decks are still full of expensive "proprietary" claims.

Why?

Because it signals seriousness to investors who aren't always digging beneath the buzzwords.

The result is predictable:

Startups with $50M raised... and no users.

Products slower, clumsier, and less reliable than basic wrappers on public models.

Features that cost millions of dollars and a team of engineers performing WORSE than foundational GPT!

Capital incinerated in the name of defensibility that doesn't exist.

When Proprietary Actually Matters

This isn’t to say that proprietary technology is always a mistake.

Strategic proprietary investment can create real moats — but only under specific conditions:

Security and Compliance: Certain industries (healthcare, finance, government) require full control over model behaviour and data pathways. Privacy concerns sometimes mandate private models.

Deep Domain Specialisation: If the problem space is extremely narrow — medical diagnostics, niche legal frameworks, scientific discovery — public models may simply not perform at acceptable levels.

Unique Data Advantage: If a startup has proprietary access to data no one else can replicate, training a model can yield real, defensible differentiation.

Architectural Control for Scale: In very large systems, cost optimisation or latency demands may make foundational model reliance unsustainable long-term.

But outside these edge cases, most startups should ask themselves: Is building our own model creating differentiated user value, or just burning VC cash?

Often, the honest answer is uncomfortable.

In quite a lot of case, which is what we are doing, a combined approach makes sense. Find niche areas you can add value by creating a model, for specific datasets that are too small for foundational models to go deeply into - but for a lot of use cases you simply do not need to fix whats not broken.

Wrappers Shouldn't Be An Insult

Dismissive phrases like "they're just a wrapper" miss the point entirely. Lots of good business are 'wrappers'. What is Deliveroo apart from a wrapper for delivering an existing product not made by them? What about a marketplace like Amazon? Businesses using AI can be an interface that delivers AI - uniquely packaged, piping hot, and just the way the customer ordered.

Foundational models are now infrastructure, like cloud servers became 15 years ago.

No one today mocks a SaaS company for not building their own data centers.

In the AI economy, the same pattern is emerging:

The value is shifting to how you orchestrate, specialise, and deliver intelligence, not how many GPUs you burned last quarter.

Smart "wrappers" build durable companies by:

• Relentlessly focusing on specific user workflows - eg: not "text generation" — but "contract risk flagging for mid-market M&A teams under UK law."

• Owning the user interface and experience - UX as a moat: intuitive, fast, trusted, embedded.

• Embedding domain expertise into prompts, workflows, and outputs- AI that understands your world, not just words.

• Engineering smarter prompts, chains, and validation layers - Not single prompts — systems of prompts that optimise outcomes.

Done right, wrappers become full-stack solutions.They stop being "just a UI on GPT" and become category leaders in their verticals.

The User Doesn't Care About Your Prop IP!

Owning the best model won't matter if you don't own the user.

We've seen this play out before:

AWS commoditized infrastructure, but the winners were those who built the best applications on top.

Mobile apps exploded when companies understood distribution — not because they re-invented TCP/IP.

CRM platforms like Salesforce dominate because they embedded into workflows, not because they built their own database engines.

In AI, the pattern will be the same:

Those who own distribution, workflow integration, user trust, and UX will build defensible businesses.

Foundational models are racing toward commoditisation.

Distribution, usability, insight — those are the new strategic high grounds.

Real-World Proof: Smart Wrappers Winning Today

This isn't theoretical. It's already happening.

Jasper.ai turned early GPT-3 access into a ~$1.5B company by specializing in marketing copy workflows — long before many even understood what prompting was.

Harvey.ai is building vertical dominance in legal AI, starting by leveraging OpenAI models — not building new ones.

Character.ai created a billion-dollar social AI experience with nothing but clever prompt engineering, UX design, and user understanding.

None of them needed to burn $100M training new models from scratch to create defensible value.

Why I Don't Care About Custom Models

If an investor thinks we are uncompetitive because of our lack of investment in proprietary models, they probably don't get AI - and how to actually find value in a world of giants like Open AI. There is IP in user interface, your multi layered prompts, your system design and APIs. You are creating IP - just not the type which requires huge datasets and GPUs. Building a custom model is not the only way to create a moat and advantage.

Investors that understand where I am going with this article , will see it as a strategic advantage to have a company that understands ROI, where to invest in R&D and where to invest in user experience. They will see that as a frugal, sensible approach rather than a disadvantage.

We make every dollar work harder by asking ruthlessly strategic questions:

Where does building our own tech create true defensibility?

Where can we leverage existing infrastructure smarter and faster?

How do we design interfaces that turn intelligence into adoption, not just outputs into noise?

How do we understand and own specific user problems so deeply that switching away becomes painful?

Proprietary IP is a tool, not a trophy.

We invest in it when it makes sense and when it give us a true unique advantage — and avoid it when it’s vanity.

In a world obsessed with burn rates, we’re obsessed with return rates.

That’s how we make capital — and capabilities — go further.

Also I know our true advantage lies in my unique mix of understanding fashion deeply, design deeply, business tech and AI. A prop model can't beat a system using a foundational model designed by someone who understands the customer and product. It will just spit out stuff - not stuff people like.

Final Thought: Build Strategically, Not Performatively

The next wave of AI winners won’t be determined by who raised the most to train a model.

It will be determined by who thought hardest about what to build, why to build, and when not to build.

Proprietary is powerful — but strategy is everything.

Execution and design, not just engineering, will define the future. As it always has.

AI has changed a lot but not changed the fundamentals - its still all about the user, the problem, the solution. Nothing else.